spark-on-k8s

README

This is not an officially supported Google product.

Community

• Slack channel.

Project Status

Project status: alpha

Spark Operator is still under active development. Backward compatibility of the APIs is not guaranteed for alpha releases.

Customization of Spark pods, e.g., mounting arbitrary volumes and setting pod affinity, is currently experimental and implemented using a Kubernetes Mutating Admission Webhook, which became beta in Kubernetes 1.9. The mutating admission webhook is disabled by default but can be enabled if there are needs for pod customization. Check out the Quick Start Guide on how to enable the webhook.

Prerequisites

• Version >= 1.8 of Kubernetes.
• Version >= 1.9 of Kubernetes if using the mutating admission webhook for Spark pod customization.

Spark Operator relies on garbage collection support for custom resources that is available in Kubernetes 1.8+ and optionally the Mutating Admission Webhook which is available in Kubernetes 1.9+.

Due to this bug in Kubernetes 1.9 and earlier, CRD objects with escaped quotes (e.g., spark.ui.port\") in map keys can cause serialization problems in the API server. So please pay extra attention to make sure no offending escaping is in your SparkAppliction CRD objects, particularly if you use Kubernetes prior to 1.10.

Get Started

Get started quickly with the Spark Operator using the Quick Start Guide.

If you are running the Spark Operator on Google Kubernetes Engine and want to use Google Cloud Storage (GCS) and/or BigQuery for reading/writing data, also refer to the GCP guide.

For more information, check the Design, API Specification and detailed User Guide.

Overview

Spark Operator aims to make specifying and running Spark applications as easy and idiomatic as running other workloads on Kubernetes. It uses Kubernetes custom resources for specifying, running, and surfacing status of Spark applications. For a complete reference of the custom resource definitions, please refer to the API Definition. For details on its design, please refer to the design doc. It requires Spark 2.3 and above that supports Kubernetes as a native scheduler backend.

Spark Operator currently supports the following list of features:

• Supports Spark 2.3 and up.
• Enables declarative application specification and management of applications through custom resources.
• Automatically runs spark-submit on behalf of users for each SparkApplication eligible for submission.
• Provides native cron support for running scheduled applications.
• Supports customization of Spark pods beyond what Spark natively is able to do through the mutating admission webhook, e.g., mounting ConfigMaps and volumes, and setting pod affinity/anti-affinity.
• Supports automatic application re-submission for updated SparkAppliation objects with updated specification.
• Supports automatic application restart with a configurable restart policy.
• Supports automatic retries of failed submissions with optional linear back-off.
• Supports mounting local Hadoop configuration as a Kubernetes ConfigMap automatically via sparkctl.
• Supports automatically staging local application dependencies to Google Cloud Storage (GCS) via sparkctl.

Motivations

This approach is completely different than the one that has the submission client creates a CRD object. Having externally created and managed CRD objects offer the following benefits:

• Things like creating namespaces and setting up RBAC roles and resource quotas represent a separate concern and are better done before applications get submitted.
• With the external CRD controller submitting applications on behalf of users, they don't need to deal with the submission process and the spark-submit command. Instead, the focus is shifted from thinking about commands to thinking about declarative YAML files describing Spark applications that can be easily version controlled.
• Externally created CRD objects make it easier to integrate Spark application submission and monitoring with users' existing pipelines and tooling on Kubernetes.
• Internally created CRD objects are good for capturing and communicating application/executor status to the users, but not for driver/executor pod configuration/customization as very likely it needs external input. Such external input most likely need additional command-line options to get passed in.

Additionally, keeping the CRD implementation outside the Spark repository gives us a lot of flexibility in terms of functionality to add to the CRD controller. We also have full control over code review and release process.